A diatom small target detection method and device for drowning diagnosis

Abstract

The present application relates to a kind of diatom small target detection method and device for drowning diagnosis, the method includes: S1: collection diatom image dataset, image and format are handled;S2: utilize multilayer convolution extraction image feature, generate multi-scale feature map;S3: construct feature enhancement and selection module SFE strengthen the response of degenerative area and strengthen edge information, generate boundary enhancement feature map;S4: construct multi-scale positive and negative convolution module MPNC to local enhancement is carried out to boundary enhancement feature map, generates enhancement feature map;S5: utilize weighted fusion to carry out cross-scale information integration to different resolution enhancement feature map, and output target boundary frame and corresponding class prediction result.By the present application, it can effectively solve the problem that small target diatom is weak in feature expression in detection, and detection effect is poor under complex background.The present application is as a kind of small target diatom detection method and device based on multi-scale feature perception, can be widely applied in microscopic image processing, classification and other fields.

Description

Technical Field

[0001] This invention relates to the field of target detection, and more specifically to a method and apparatus for detecting small diatom targets for drowning diagnosis. Background Technology

[0002] Diatom detection is an important diagnostic method in drowning diagnosis. It can be used to determine whether a body retrieved from water is due to drowning or post-mortem disposal, and can even infer the point of entry into the water based on diatom species. The types and quantities of diatoms are affected by water quality; different water conditions will produce different types and quantities of diatoms. Therefore, diatom detection provides a reliable diagnostic basis for drowning cases. Research has found that diatoms, being small targets, are prone to being missed due to low image resolution, indistinct features, and susceptibility to occlusion. Designing an accurate and efficient model is a challenge. Therefore, improving the accuracy of small target detection in complex backgrounds is crucial.

[0003] There are roughly four main types of existing small object detection methods. First, multi-scale feature fusion methods rely on high-resolution feature maps for small object detection, but the fusion process may dilute or weaken fine-grained features from shallow layers, reducing the accuracy of small object detection. Second, context-aware enhancement methods, while providing more background information, are easily interfered with by complex backgrounds, causing small object features to be obscured. Third, data augmentation methods can only supplement the distribution of training data and cannot fundamentally solve the problem of insufficient small object feature extraction capability from the model structure. Fourth, special convolutional designs, while reducing computational cost, have limited expressive power, especially in terms of sensitivity to capturing fine-grained features of small objects.

[0004] In summary, existing technologies for small target detection suffer from problems such as multi-scale targets and weak feature representation. Therefore, how to effectively capture and utilize the subtle features of small targets to improve the detection accuracy of the model has become a key issue that urgently needs to be addressed in current research. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and apparatus for detecting small diatom targets in drowning diagnosis, which can effectively solve the problems of weak feature expression and poor detection results in complex backgrounds when detecting small diatom targets.

[0006] Specifically, the method includes the following steps:

[0007] S1: Collect diatom detection image dataset and process the images and their formats;

[0008] S2: Extract image features using multi-layer convolution to generate multi-scale feature maps;

[0009] S3: Construct the Feature Enhancement and Selection Module (SFE) to enhance the response of degraded regions and enhance edge information, generating boundary enhanced feature maps;

[0010] S4: Construct a multi-scale positive-negative convolution module (MPNC) to locally enhance the boundary enhancement feature map and generate enhanced features;

[0011] S5: Use weighted fusion to integrate cross-scale information from enhanced feature maps of different resolutions, and output the target bounding box and the corresponding category prediction results.

[0012] Preferably, S1 includes the following steps:

[0013] S1.1: Collect image datasets of various types of diatoms and use annotation tools to accurately label the targets;

[0014] S1.2: Perform uniform size adjustment on all images and normalize the image pixel values;

[0015] S1.3: Convert the annotation file to .txt format, including the center coordinates of the target ( x, y ), width and height w, h ), rotation angle ( theta ).

[0016] Preferably, S2 includes the following steps:

[0017] S2.1: Perform feature extraction and fusion on the image;

[0018] Preferably, S2.1 includes the following steps:

[0019] S2.1.1: Input Image Preliminary features were extracted using multiple YOLO11 CBS modules (Convolution BatchNorm-SiLU). The expression is as follows:

[0020]

[0021] ,

[0022] In the above formula, This represents the SiLU activation function. Indicates batch normalization. This represents a two-dimensional convolution operation;

[0023] S2.1.2: Preliminary features are obtained by fusing deep and shallow features through YOLO11's multi-scale feature extraction module C3K2 (Cross-Stage Partial Kernel2). The expression is as follows:

[0024] ,

[0025]

[0026] ,

[0027] In the above formula, This represents the segmentation operation based on the input features. and yes The segmented features and the sum of the number of channels equals The number of channels, Indicates that the convolution kernel is n Convolution operation, This indicates a splicing operation. Indicates stacking n A C3 module, in which a main branch directly skips connections and a sub-branch includes convolutions and Bottleneck structures;

[0028] S2.1.3: High-level features are obtained through YOLO11's long-range modeling module C2PSA (CSP + PSABlock block) using a global attention mechanism. The expression is as follows:

[0029] ,

[0030]

[0031] ,

[0032] In the above formula, and yes The segmented features and the sum of the number of channels equals The number of channels, It was through n The result of stacking PSABlock blocks, each PSABlock block includes Attention, FNN and convolutional modules;

[0033] S2.1.4: Global features are obtained by aggregating contextual information at different scales through the Spatial Pyramid Pooling-Fast (SPPF) module. The expression is as follows:

[0034]

[0035] ,

[0036] In the above formula, Indicates the pooling window size is n×n Max pooling operation.

[0037] S2.2: Output set of multi-layer features from the YOLO11 backbone Extract multi-scale features.

[0038] Preferably, S3 includes the following steps:

[0039] S3.1: Construct the Feature Enhancement and Selection Module (SFE).

[0040] S3.2: Extract global features of degraded regions from the multi-scale feature maps output by the YOLO11 backbone using pooling and convolution operations. (i=2,3,4);

[0041] Preferably, S3.2 includes the following steps:

[0042] S3.2.1: Extract global features of degenerate regions from multi-scale feature maps using pooling operations, as shown in the following expression:

[0043] ,

[0044] In the above formula, These represent feature maps at multiple scales from the output of the backbone network. The feature map represents multiple scales after the pooling operation.

[0045] S3.2.2: Extract channel-level features using 1×1 convolution, and model the degenerate region features in the multi-scale feature map channel by channel. The expression is as follows:

[0046] ,

[0047] ,

[0048] In the above formula, This represents a depthwise convolution operation. This indicates the characteristics of the degraded region.

[0049] S3.2.3: Multiply and weight the degraded region features in the multi-scale feature map with the original feature map point by point to generate features. The expression is as follows:

[0050] ,

[0051] In the above formula ， This indicates point-by-point multiplication.

[0052] S3.3: Utilize Discrete Wavelet Transform (DWT) to enhance the edge information of global features, and weightedly fuse global features and enhanced edge features to generate boundary enhancement features. ;

[0053] Preferably, S3.3 includes the following steps:

[0054] S3.3.1: Using Discrete Wavelet Transform (DWT) to transform features It is decomposed into a low-frequency component and three high-frequency components, namely a horizontal high-frequency component, a vertical high-frequency component, and a diagonal high-frequency component, as expressed below:

[0055] ,

[0056] In the above formula, This represents the discrete wavelet transform operation. The low-frequency components representing features at each scale. These represent the horizontal, vertical, and diagonal components of the high-frequency components of features at each scale, respectively.

[0057] S3.3.2: The three high-frequency components are fused and enhanced, while the low-frequency components are slightly suppressed. The high-frequency and low-frequency features are then refused to generate the new features. The expression is as follows:

[0058] ,

[0059] ,

[0060] ,

[0061] ,

[0062] In the above formula, This indicates the characteristics after re-fusion.

[0063] S3.3.3: Features Upsampled to the size of the original feature map, the output is then weighted and fused. and Generate boundary enhancement features The expression is as follows:

[0064] ,

[0065] ,

[0066] In the above formula, Indicates an upsampling operation. and These are learnable weight parameters used to balance the weights of degraded regions and edge information.

[0067] Preferably, S4 includes the following steps:

[0068] S4.1: Construct a multi-scale positive-negative convolution module (MPNC).

[0069] S4.2: Utilize a multi-branch convolutional structure to extract multi-scale and multi-directional features, with each branch outputting features. ;

[0070] Preferably, S4.2 includes the following steps:

[0071] S4.2.1: Branch 1 extracts global features between channels through 1×1 convolution, as shown in the following expression:

[0072] ,

[0073] S4.2.2: Branches 2 and 3 successively use convolutions along different paths (1×3 and 3×1) to capture horizontal, vertical, and local contextual information, and combine this with deconvolution to improve resolution. The expression is as follows:

[0074] ,

[0075] ,

[0076] S4.2.3: Branch 4 directly extracts high-resolution features through 1×1 convolution + deconvolution + 3×3 convolution, as shown in the following expression:

[0077] ,

[0078] In the above formula, This indicates the deconvolution operation.

[0079] S4.3: Merge the enhanced features of all branches to generate enhanced features, as shown in the following expression;

[0080] ,

[0081] In the above formula, This is the final output feature of the MPNC module.

[0082] Preferably, S5 includes the following steps:

[0083] S5.1: Integrate information from enhanced feature maps of different resolutions;

[0084] Preferably, S5.1 includes the following steps:

[0085] S5.1.1: By upsampling the high-level feature maps and downsampling the low-level feature maps, the feature map sizes are aligned.

[0086] S5.1.2: Integrate information from different scales through cross-scale feature fusion;

[0087] S5.2: Extract multi-scale information from the fused features and output multi-scale features.

[0088] Preferably, S5.2 includes the following steps:

[0089] Multi-scale features are further extracted using standard convolution and dilated convolution. High-resolution, medium-resolution, and low-resolution features are then uniformly integrated and fed into the detection head for task processing. The expression is as follows:

[0090] ,

[0091] In the above formula, These are represented as the integrated high, medium, and low resolution feature maps, respectively. These are represented as feature maps of different scales fed into the Head network.

[0092] S5.3: Perform classification prediction and bounding box parameter prediction on feature maps of different scales for each input;

[0093] Preferably, S5.3 includes the following steps:

[0094] S5.3.1: Using the CBS module and Multi-Scale Convolutional Decomposition (MSCD), the expression is as follows:

[0095] ,

[0096]

[0097] In the above formula, This indicates the output of the CBS module. This indicates the output of the MSDC module.

[0098] S5.3.2: Use independent convolutional layers to predict the target category for the features at each location, as shown in the following expression:

[0099] ,

[0100] In the above formula, Indicates the number of categories. This represents the classification prediction parameters.

[0101] S5.3.3: Using another independent convolutional layer, predict the regression parameters of the bounding box, as shown in the following expression:

[0102] ,

[0103] In the above formula, This represents the predicted bounding box parameters.

[0104] S5.4: Integrate the classification and bounding box prediction results to generate the final detection result.

[0105] Preferably, S6.2 includes the following steps:

[0106] Decode the bounding boxes, converting the offsets into actual coordinates in the image. Use NMS to remove redundant bounding boxes, retaining the target boxes with the highest confidence. The expression is as follows:

[0107] ,

[0108] ,

[0109] In the above formula, This indicates nonmaximum suppression.

[0110] This invention first extracts image features using multi-layer convolution to generate multi-scale feature maps. It then enhances the response of degraded regions by constructing a Feature Enhancement and Selection (SFE) module and extracts edge information using Discrete Wavelet Transform (DWT). Furthermore, it locally enhances the feature maps by constructing a Multi-Scale Convolutional (MPNC) module. Finally, it integrates cross-scale information from feature maps of different resolutions using weighted fusion, outputting accurate target bounding boxes and corresponding category predictions. Simultaneously, this invention effectively addresses the challenges of existing neural network models in small target detection. On one hand, the Feature Enhancement and Selection module highlights degraded regions, effectively strengthening edge information; on the other hand, the multi-branch convolutional structure effectively increases local information and suppresses background. Furthermore, the CBS module and Multi-Scale Decomposed Convolution (MSCD) enhance the ability to extract fine-grained features from small targets, ultimately improving the detection accuracy of small target diatom images. Attached Figure Description

[0111] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following is a brief introduction to the drawings used in the prior art and embodiments. The following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0112] Figure 1 is a flowchart illustrating a method for detecting small diatom targets for drowning diagnosis according to the present invention.

[0113] Figure 2 is a schematic diagram of a specific embodiment of the diatom small target detection method for drowning diagnosis according to the present invention;

[0114] Figure 3 This is a schematic diagram of the SFE module structure of a diatom small target detection device for drowning diagnosis according to the present invention;

[0115] Figure 4 This is a schematic diagram of the MPCN module structure of a diatom small target detection device for drowning diagnosis according to the present invention; Detailed Implementation Plan

[0116] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0117] This invention provides a method and apparatus for detecting small diatom targets for drowning diagnosis, which improves the fine-grained feature extraction capability of small targets at the algorithm level, thereby improving the detection accuracy of small target diatom images.

[0118] A typical embodiment of the present invention, taking a diatom image under a microscope magnified 1800x as an example, refers to... Figure 1 , Figure 2 The method includes the following steps:

[0119] S1: Collect diatom detection image dataset and process the images and their formats;

[0120] S2: Extract image features using multi-layer convolution to generate multi-scale feature maps;

[0121] S3: Construct the Feature Enhancement and Selection Module (SFE) to enhance the response of degraded regions and enhance edge information, generating boundary enhanced feature maps;

[0122] S4: Construct a multi-scale positive-negative convolution module (MPNC) to locally enhance the boundary enhancement feature map and generate enhanced features;

[0123] S5: Use weighted fusion to integrate cross-scale information from enhanced feature maps of different resolutions, and output the target bounding box and the corresponding category prediction results.

[0124] Furthermore, S1 includes the following steps:

[0125] S1.1: Collect image datasets of various types of diatoms and use annotation tools to accurately label the targets;

[0126] S1.2: Perform uniform size adjustment on all images and normalize the image pixel values;

[0127] S1.3: Convert the annotation file to .txt format, including the center coordinates of the target ( x, y ), width and height w, h ), rotation angle ( theta ).

[0128] Furthermore, S2 includes the following steps:

[0129] S2.1: Perform feature extraction and fusion on the image;

[0130] Furthermore, S2.1 includes the following steps:

[0131] S2.1.1: Input Image Preliminary features were extracted using multiple YOLO11 CBS modules (Convolution BatchNorm-SiLU). The expression is as follows:

[0132]

[0133] ,

[0134] In the above formula, This represents the SiLU activation function. Indicates batch normalization. This represents a two-dimensional convolution operation;

[0135] S2.1.2: Preliminary features are obtained by fusing deep and shallow features through YOLO11's multi-scale feature extraction module C3K2 (Cross-Stage Partial Kernel2). The expression is as follows:

[0136] ,

[0137]

[0138] ,

[0139] In the above formula, This represents the segmentation operation based on the input features. and yes The segmented features and the sum of the number of channels equals The number of channels, Indicates that the convolution kernel is n Convolution operation, This indicates a splicing operation. Indicates stacking n A C3 module, in which a main branch directly skips connections and a sub-branch includes convolutions and Bottleneck structures;

[0140] S2.1.3: High-level features are obtained through YOLO11's long-range modeling module C2PSA (CSP + PSABlock block) using a global attention mechanism. The expression is as follows:

[0141] ,

[0142]

[0143] ,

[0144] In the above formula, and yes The segmented features and the sum of the number of channels equals The number of channels, It was through n The result of stacking PSABlock blocks, each PSABlock block includes Attention, FNN and convolutional modules;

[0145] S2.1.4: Global features are obtained by aggregating contextual information at different scales through the Spatial Pyramid Pooling-Fast (SPPF) module. The expression is as follows:

[0146]

[0147] ,

[0148] In the above formula, Indicates the pooling window size is n×n Max pooling operation.

[0149] S2.2: Output set of multi-layer features from the YOLO11 backbone Extract multi-scale features.

[0150] Furthermore, refer to Figure 3 S3 includes the following steps:

[0151] S3.1: Construct the Feature Enhancement and Selection Module (SFE).

[0152] S3.2: Extract global features of degraded regions from the feature maps output by the YOLO11 backbone using pooling and convolution operations. (i=2,3,4);

[0153] Furthermore, S3.2 includes the following steps:

[0154] S3.2.1: Extract global features of degenerate regions from multi-scale feature maps using pooling operations, as shown in the following expression:

[0155] ,

[0156] In the above formula, These represent feature maps at multiple scales from the output of the backbone network. The feature map represents multiple scales after the pooling operation.

[0157] S3.2.2: Extract channel-level features using 1×1 convolution, and model the degenerate region features in the feature map channel by channel. The expression is as follows:

[0158] ,

[0159] ,

[0160] In the above formula, This represents a depthwise convolution operation. This indicates the characteristics of the degraded region.

[0161] S3.2.3: Multiply and weight the degraded region features in the multi-scale feature map with the original feature map point by point to generate features. The expression is as follows:

[0162] ,

[0163] In the above formula ， This indicates point-by-point multiplication.

[0164] S3.3: Utilize Discrete Wavelet Transform (DWT) to enhance the edge information of global features, and weightedly fuse global features and enhanced edge features to generate boundary enhancement features. ;

[0165] Furthermore, S3.3 includes the following steps:

[0166] S3.3.1: Using Discrete Wavelet Transform (DWT) to transform features It is decomposed into a low-frequency component and three high-frequency components, namely a horizontal high-frequency component, a vertical high-frequency component, and a diagonal high-frequency component, as expressed below:

[0167] ,

[0168] In the above formula, This represents the discrete wavelet transform operation. The low-frequency components representing features at each scale. These represent the horizontal, vertical, and diagonal components of the high-frequency components of features at each scale, respectively.

[0169] S3.3.2: The three high-frequency components are fused and enhanced, while the low-frequency components are slightly suppressed. The high-frequency and low-frequency features are then refused to generate the new features. The expression is as follows:

[0170] ,

[0171] ,

[0172] ,

[0173] ,

[0174] In the above formula, This indicates the characteristics after re-fusion.

[0175] S3.3.3: Features Upsampled to the size of the original feature map, the output is then weighted and fused. and Generate boundary enhancement features The expression is as follows:

[0176] ,

[0177] ,

[0178] In the above formula, Indicates an upsampling operation. and These are learnable weight parameters used to balance the weights of degraded regions and edge information.

[0179] Furthermore, refer to Figure 4 S4 includes the following steps:

[0180] S4.1: Construct a multi-scale positive-negative convolution module (MPNC).

[0181] S4.2: Utilize a multi-branch convolutional structure to extract multi-scale and multi-directional features, with each branch outputting features. ;

[0182] Furthermore, S4.2 includes the following steps:

[0183] S4.2.1: Branch 1 extracts global features between channels through 1×1 convolution, as shown in the following expression:

[0184] ,

[0185] S4.2.2: Branches 2 and 3 successively use convolutions along different paths (1×3 and 3×1) to capture horizontal, vertical, and local contextual information, and combine this with deconvolution to improve resolution. The expression is as follows:

[0186] ,

[0187] ,

[0188] S4.2.3: Branch 4 directly extracts high-resolution features through 1×1 convolution + deconvolution + 3×3 convolution, as shown in the following expression:

[0189] ,

[0190] In the above formula, This indicates the deconvolution operation.

[0191] S4.3: Merge the enhanced features of all branches to generate enhanced features, as shown in the following expression;

[0192] ,

[0193] In the above formula, This is the final output feature of the MPNC module.

[0194] Furthermore, S5 includes the following steps:

[0195] S5.1: Integrate information from enhanced feature maps of different resolutions;

[0196] Furthermore, S5.1 includes the following steps:

[0197] S5.1.1: By upsampling the high-level feature maps and downsampling the low-level feature maps, the feature map sizes are aligned.

[0198] S5.1.2: Integrate information from different scales through cross-scale feature fusion;

[0199] S5.2: Extract multi-scale information from the fused features and output multi-scale features.

[0200] Furthermore, S5.2 includes the following steps:

[0201] Multi-scale features are further extracted using standard convolution and dilated convolution. High-resolution, medium-resolution, and low-resolution features are then uniformly integrated and fed into the detection head for task processing. The expression is as follows:

[0202] ,

[0203] In the above formula, These are represented as the integrated high, medium, and low resolution feature maps, respectively. These are represented as feature maps of different scales fed into the Head network.

[0204] S5.3: Perform classification prediction and bounding box parameter prediction on feature maps of different scales for each input;

[0205] Furthermore, S5.3 includes the following steps:

[0206] S5.3.1: Using the CBS module and Multi-Scale Convolutional Decomposition (MSCD), the expression is as follows:

[0207] ,

[0208]

[0209] In the above formula, This indicates the output of the CBS module. This indicates the output of the MSDC module.

[0210] S5.3.2: Use independent convolutional layers to predict the target category for the features at each location, as shown in the following expression:

[0211] ,

[0212] In the above formula, Indicates the number of categories. This represents the classification prediction parameters.

[0213] S5.3.3: Using another independent convolutional layer, predict the regression parameters of the bounding box, as shown in the following expression:

[0214] ,

[0215] In the above formula, This represents the predicted bounding box parameters.

[0216] S5.4: Integrate the classification and bounding box prediction results to generate the final detection result.

[0217] Furthermore, S5.4 includes the following steps:

[0218] Decode the bounding boxes, converting the offsets into actual coordinates in the image. Use NMS to remove redundant bounding boxes, retaining the target boxes with the highest confidence. The expression is as follows:

[0219] ,

[0220] ,

[0221] In the above formula, This indicates nonmaximum suppression.

[0222] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make other equivalent modifications or substitutions without departing from the spirit of the invention. These equivalent modifications or substitutions are included within the scope defined by the claims.

Claims

1. A method for detecting small diatom targets for drowning diagnosis, characterized in that, The method includes the following steps: S1: Collect diatom detection image dataset and process the images and their formats; S2: Extract image features using multi-layer convolution to generate multi-scale feature maps; S3: Construct the Feature Enhancement and Selection (SFE) module to enhance the response of the degraded region and enhance the edge information, generating a boundary enhanced feature map; S3 includes the following steps: S3.1: Construct the Feature Enhancement and Selection Module (SFE). S3.2: Extract global features of degenerate regions from the multi-scale feature maps output by the YOLO11 backbone using pooling and convolution operations. i=2,3,4; S3.3: Utilize Discrete Wavelet Transform (DWT) to enhance the edge information of global features, and weightedly fuse global features and enhanced edge features to generate boundary enhancement features. ; S4: Construct a multi-scale positive-negative convolution module (MPNC) to locally enhance the boundary enhancement feature map and generate an enhanced feature map; S5: Use weighted fusion to integrate cross-scale information from enhanced feature maps of different resolutions and output target bounding boxes and corresponding category prediction results.

2. The method for detecting small diatom targets for drowning diagnosis according to claim 1, characterized in that, S1 includes the following steps: S1.1: Collect image datasets of various types of diatoms and use annotation tools to accurately label the targets; S1.2: Perform uniform size adjustment on all images and normalize the image pixel values; S1.3: Convert the annotation file to .txt format, including the center coordinates of the target ( x, y ), width and height w, h ), rotation angle ( theta ).

3. The method for detecting small diatom targets for drowning diagnosis according to claim 1, characterized in that, S2 includes the following steps: S2.1: Perform feature extraction and fusion on the image; S2.1 includes the following steps: S2.1.1: Input Image Preliminary features were extracted using multiple YOLO11 CBS modules (Convolution BatchNorm-SiLU). The expression is as follows: , In the above formula, This represents the SiLU activation function. Indicates batch normalization. Represents a two-dimensional convolution operation; S2.1.2: Preliminary features are obtained by fusing deep and shallow features through YOLO11's multi-scale feature extraction module C3K2 (Cross-Stage Partial Kernel2). The expression is as follows: , , In the above formula, This represents the segmentation operation based on the input features. and yes The segmented features and the sum of the number of channels equals The number of channels, Indicates that the convolution kernel is n Convolution operation, This indicates a splicing operation. Indicates stacking n A C3 module, in which a main branch directly skips connections and a sub-branch includes convolutions and Bottleneck structures; S2.1.3: High-level features are obtained through YOLO11's long-range modeling module C2PSA (CSP + PSABlock block) using a global attention mechanism. The expression is as follows: , , In the above formula, and yes The segmented features and the sum of the number of channels equals The number of channels, It was through n The result of stacking PSABlock blocks, each PSABlock block includes Attention, FNN and convolutional modules; S2.1.4: Global features are obtained by aggregating contextual information at different scales through the Spatial Pyramid Pooling-Fast (SPPF) module. The expression is as follows: , In the above formula, Indicates the pooling window size is n×n max pooling operation; S2.2: Output set of multi-layer features from the YOLO11 backbone Extract multi-scale features.

4. The method for detecting small diatom targets for drowning diagnosis according to claim 1, characterized in that, S3 includes the following steps: S3.2 includes the following steps: S3.2.1: Extract global features of degenerate regions from multi-scale feature maps using pooling operations, as shown in the following expression: , In the above formula, These represent feature maps at multiple scales from the output of the YOLO11 backbone. Feature maps representing multiple scales after pooling operations; S3.2.2: Extract channel-level features using 1×1 convolution, and model the degenerate region features in the multi-scale feature map channel by channel. The expression is as follows: , , In the above formula, This represents a depthwise convolution operation. Indicates the characteristics of the degraded region; S3.2.3: Multiply and weight the degraded region features in the multi-scale feature map with the original feature map point by point to generate features. The expression is as follows: , In the above formula ， This indicates point-by-point multiplication. S3.3 includes the following steps: S3.3.1: Using Discrete Wavelet Transform (DWT) to transform features It is decomposed into a low-frequency component and three high-frequency components, namely a horizontal high-frequency component, a vertical high-frequency component, and a diagonal high-frequency component, as expressed below: , In the above formula, This represents the discrete wavelet transform operation. The low-frequency components representing features at each scale. These represent the horizontal, vertical, and diagonal components of the high-frequency components of features at each scale, respectively. S3.3.2: The three high-frequency components are fused and enhanced, while the low-frequency components are slightly suppressed. The high-frequency and low-frequency features are then refused to generate the new features. The expression is as follows: , , , , In the above formula, Indicates the characteristics after re-fusion; S3.3.3: Features Upsampled to the size of the original feature map, the output is then weighted and fused. and Generate boundary enhancement features The expression is as follows: , , In the above formula, Indicates an upsampling operation. and These are learnable weight parameters used to balance the weights of degraded regions and edge information.

5. The method for detecting small diatom targets for drowning diagnosis according to claim 1, characterized in that, S4 includes the following steps: S4.1: Construct a multi-scale positive-negative convolution module (MPNC). S4.2: Utilize a multi-branch convolutional structure to extract multi-scale and multi-directional features, with each branch outputting features. ; S4.2 includes the following steps: S4.2.1: Branch 1 extracts global features between channels through 1×1 convolution, as shown in the following expression: , S4.2.2: Branches 2 and 3 successively capture horizontal, vertical, and local contextual information through 1×3 and 3×1 convolutions along different paths, and combine this with deconvolution to improve resolution. The expression is as follows: , , S4.2.3: Branch 4 directly extracts high-resolution features through 1×1 convolution + deconvolution + 3×3 convolution, as shown in the following expression: , In the above formula, Indicates the deconvolution operation; S4.3: Merge the enhanced features of all branches to generate enhanced features, as shown in the following expression; , In the above formula, This is the final output feature of the MPNC module.

6. The method for detecting small diatom targets for drowning diagnosis according to claim 1, characterized in that, S5 includes the following steps: S5.1: Integrate information from enhanced feature maps of different resolutions; S5.1 includes the following steps: S5.1.1: By upsampling the high-level feature maps and downsampling the low-level feature maps, the feature map sizes are aligned. S5.1.2: Integrate information from different scales through cross-scale feature fusion; S5.2: Extract multi-scale information from the fused features and output multi-scale features; S5.2 includes the following steps: Multi-scale features are further extracted using standard convolution and dilated convolution. High-resolution, medium-resolution, and low-resolution features are then uniformly integrated and fed into the detection head for task processing. The expression is as follows: , In the above formula, These are represented as the integrated high, medium, and low resolution feature maps, respectively. These are represented as feature maps of different scales fed into the Head network in YOLO11. S5.3: Perform classification prediction and bounding box parameter prediction on feature maps of different scales for each input; S5.3 includes the following steps: S5.3.1: Using the CBS module and Multi-Scale Convolutional Decomposition (MSCD), the expression is as follows: , In the above formula, This indicates the output of the CBS module. Indicates the output of the MSDC module; S5.3.2: Use independent convolutional layers to predict the target category for the features at each location, as shown in the following expression: , In the above formula, Indicates the number of categories. Indicates the classification prediction parameters; S5.3.3: Using another independent convolutional layer, predict the regression parameters of the bounding box, as shown in the following expression: , In the above formula, Indicates the parameters of the predicted bounding box; S5.4: Integrate the classification and bounding box prediction results to generate the final detection results; S5.4 includes the following steps: Decode the bounding boxes, converting the offsets into actual coordinates in the image. Use NMS to remove redundant bounding boxes, retaining the target boxes with the highest confidence. The expression is as follows: , , In the above formula, This indicates nonmaximum suppression.

Images